A cardiac imaging program uses flexday.ai where a millimetre is the whole specification

Stent positioning and placement sit at the far end of the accuracy scale, built on complex segmentation where a millimetre of drift is the difference between a correct and an incorrect reading, so the evaluation bar is set accordingly.

Golden sets are owned by the clinicians themselves, not by the engineering team, regression runs on every release before it ships, and a human clinician makes the final call on every case. The agent's job is to narrow what that clinician has to check, not to replace the check.